nnabla-ggnn: NNabla Implementation of GG-NN

This repository is a NNabla implementation of Gated Graph Sequence Neural Networks (GG-NN) proposed in the paper Gated Graph Sequence Neural Networks by Y.Li, D.Tarlwo, M.Brockschmdit, and R. Zemel. GG-NNs can use graph-structured data as inputs of neural networks, and gets high accuracy on some bAbI-tasks. This implementation is tested with bAbI 15 and bAbI 19, and gets high accuracy (100% for bAbI15 and 95% for bAbI 19). The official implementation is available in the GitHub repository.

Requirements

• Python 3.x (tested with Python 3.6.5)
• NNabla 0.9.9

Run Examples

bAbI 15

$ babi-tasks 15 1000 > train.txt # Notes: babi-tasks can be installed from https://github.com/facebook/bAbI-tasks
$ babi-tasks 15 1000 > vaild.txt
$ python ./main.py bAbI15 --train-file train.txt --valid-file valid.txt # --context cudnn

My result: get 100% validation accuracy after 200 iterations (1 epochs).

bAbI 19

$ babi-tasks 19 1000 > train.txt # Notes: babi-tasks can be installed from https://github.com/facebook/bAbI-tasks
$ babi-tasks 19 1000 > vaild.txt
$ python ./main.py bAbI10 --train-file train.txt --valid-file valid.txt # --context cudnn

My result: get 95% validation accuracy after 54000 iterations (216 epochs).

TODO

• mini-batched training
  • I didn't implement the mini-batch version.
  • I think it is not difficult to implement mini-batched training if the number of vertices in graph is same.

References

• Gated Graph Sequence Neural Networks, ICLR 2016
• the official GitHub repository
• bAbI-tasks